Reproduce “Taming Generative Synthetic Data for X-ray Prohibited Item Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jialong Sun, Hongguang Zhu, Weizhe Liu, Yunda Sun, Renshuai Tao, Yunchao Wei
Researchers developed Xsyn, a one-stage generative pipeline for X-ray prohibited item detection that uses text-grounded inpainting to create synthetic data, eliminating labor-intensive foreground preparation. The method integrates Cross-Attention Refinement for precise annotations and Background Occlusion Modeling for realism, achieving up to a 2.3% mAP increase on the PIDray dataset compared to baseline and outperforming prior synthesis methods.
Key insights
- A one-stage, text-grounded inpainting approach for X-ray synthesis effectively bypasses the labor-intensive foreground preparation inherent in previous two-stage methods.
- Cross-attention maps from diffusion models can be effectively utilized with SAM and median point sampling to achieve precise annotation refinement for generated objects, improving bounding box accuracy.
- Simulating realistic background occlusion in the latent space during the diffusion denoising process enhances the synthetic data's realism and improves detector generalization by preventing overfitting to overly clear items.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.